Multiple Metric Learning in Kernel Space for Person Re-identification

In this paper, we present multiple metric learning in kernel space to preferably get more discriminative metrics. Usually, the kernel-based approaches exploit kernel trick to map vectors that usually thousands of dimensions into high dimensional space to enhance their linear capacity. But it could loss the discriminative information in the process of projecting. To address this problem, we propose to map these feature vectors into kernel space respectively and the metrics are learned in their corresponding space. The Relief algorithm is modified here to get the weights and we can get the final result by weighting multiple metrics based on features. Experiments on the public datasets demonstrate the performance of our proposed method outperforms some state of the art methods.